How to get a YouTube transcript without timestamps.

A timestamp is not decoration added on top of the text; it is half of what a caption is. Each cue stores a start time, an end time and a line of words. SRT and WebVTT make that structure visible, and even the transcript panel on YouTube shows a time next to every row. So "give me just the words" means choosing a format that drops the other half — not editing numbers out of a file afterwards.

The short answer

Take TXT, not SRT and not VTT. On the subtitle page, paste the link, let the track load, and export TXT: the result is one line per cue with a blank line between, no cue numbers, no arrows, no timestamps. It is the only one of the three formats that is clean the moment you save it.

What you give up is small but real: with no timing you cannot jump back to the moment a line was said. If you might need that, take the SRT as well — the same track produces both, and the extra download costs nothing beyond the first read.

When the timestamps get in the way

• Word counts and reading time. 00:00:12 is not a word, but every counter that splits on whitespace counts it as one. A 40-minute video gains several hundred phantom words.
• Prompts for a language model. Timestamps are tokens the model has to read and must then decide to ignore. Ask it to summarise and you get summaries that occasionally mention "at 12 seconds"; ask it to quote and it may quote the number as part of the sentence.
• Quoting a person. A quote with 00:02:14 in front of it needs cleaning before it can go in an article, a report or a slide.
• Translation handoff. A translator working from a timestamped file spends the first pass deleting numbers instead of translating. Send clean text, keep the timed file for the person who has to sync it back.
• Search. A phrase search over clean text matches the words people actually said. Over a timed file it also matches the timestamps, which is noise.

Three different things people mean by "clean"

This is where most attempts go wrong: three requests that sound identical need three different outputs, and only one of them is actually clean.

A caption track splits on breath and on screen width, not on grammar. That is why clean text from a transcript often reads as a column of half-sentences: so what we did was / rebuild the whole intro. Joining them back into paragraphs is editing, and it is the only part of this job that cannot be automated honestly.

If all you already have is a timestamped file

Sometimes the export is not available — a colleague sent an SRT, or the only copy is a paste from a transcript panel. These three routes work, in order of how much they damage the text.

1. Re-export as TXT instead of editing the SRTDeleting the time lines from an SRT leaves the cue numbers behind, and leaves the blank lines that used to separate cues. You end up with a file that looks clean until you notice every third line is a bare digit.What to do: Take the source again and export TXT. If the source is gone, strip by pattern rather than by line position — see the next two steps.
2. Strip bracket or leading timestamps with one patternThe two shapes that cover almost everything are [00:02:14] (bracketed, any editor) and 0:12 at the start of a line (what a transcript panel paste looks like).What to do: In a text editor with regex search, replace \[\d{1,2}:\d{2}(?::\d{2})?\] with nothing, then replace ^\s*\d{1,2}:\d{2}(?::\d{2})?\s* with nothing. Collapse runs of spaces, and you are done.
3. Do not use a spreadsheet to split on the arrowSplitting 00:00:02,500 --> 00:00:05,000 on the arrow looks tidy and destroys the file: the cue number stays in one column, the text lands in the third, and re-joining them re-introduces the numbers you were trying to lose.What to do: If you must use a spreadsheet, import as plain text with no delimiter and filter rows by pattern. Otherwise use a text editor — this is a text problem, not a table problem.

How to check you actually got clean text

• No digits at the start of any line — Search for ^\d — a clean transcript should return nothing. A single leading number is usually a timestamp that survived the strip.
• No `-->` anywhere — That arrow only exists in SRT and VTT. If it survived, you are still holding a subtitle file with pieces removed.
• No `WEBVTT` on line one — The most obvious leftover, and the easiest to miss when the file was renamed rather than converted.
• Word count looks like speech — Roughly 130–160 words per minute of talking. A 10-minute video that reports 900 words lost something; one that reports 2,500 kept the timestamps.

What a clean transcript is actually for

• A prompt you can trust. No number-tokens means the model spends its context on the words, and quotes come back without a timestamp glued to the front.
• A first draft. Clean text plus the video open in another window is a faster route to an article or show notes than transcribing by ear.
• A subtitle translation handoff. The translator gets sentences; the timed file stays with whoever will re-sync afterwards.
• A searchable archive. A folder of clean transcripts is greppable. A folder of SRT files is greppable in a way that returns timestamps you did not ask for.